TimeGAN-Based Diversified Synthetic Data Generation Following BERT-Based Model for EV Battery SOC Prediction: A State-of-The-Art Approach

The recent growth of electric vehicles (EVs) usage has necessitated the development of battery modules and management systems that are both highly efficient and safe. The estimation of the state of charge (SOC) of a vehicle's battery is a critical factor that impacts the range of the vehicle and optimizes charging preferences. Numerous studies in the literature attempt to estimate it through the use of battery voltage, current, and temperature as input parameters. There are currently two significant research gaps. Firstly, the majority neglected parameters such as vehicle speed and tractive effort, which have a direct impact on battery performance. This requires a more robust model to determine the SOC, taking into account the additional dynamics of the vehicle. Secondly, there is a lack of qualitative and diverse battery datasets that are capable of predicting SOC, which is a significant limitation for practical applications. Furthermore, the procedure of acquiring data is laborious, time-consuming, and has privacy issues. In order to mitigate the aforementioned concerns, this paper presents a model based on a time-series Generative Adversarial Network (TimeGAN) that produces a diverse synthetic dataset across a wide range of ambient temperature conditions over numerous driving cycles. The objective is to improve the estimation of the SOC. Following this, a model based on Bidirectional Encoder Representations from Transformers (BERT) is suggested as a means to effectively forecast the SOC using a combined, scalable dataset that can mimic the real-world scenario. By comparing these methods to existing techniques and using a variety of evaluation metrics, their accuracy, and efficiency are validated.